Why sheet music is harder for a computer to read than text
Text recognition reads a line of letters. Music recognition has to read overlapping grids, where the meaning of each mark depends on marks above, below and far to the left.
Engineering blog
Weekly notes on findings, measurements and techniques from the StaveWave recognition pipeline.
Text recognition reads a line of letters. Music recognition has to read overlapping grids, where the meaning of each mark depends on marks above, below and far to the left.
Testing our MusicXML output in popular notation programs revealed some interesting findings. Sometimes a round trip seemed to drop slurs, ties and hairpins from our files. A small plug-in showed the import was fine, and found one real problem: overlapping slurs that share a number.
Notation-program PDFs convert to MusicXML far better than scans, so we stopped tuning the scan reader and built a front end that feeds scans to the PDF reader.
We analyzed where errors occur in our music transcription engine. The biggest share was music we read correctly and then placed at the wrong beat.